Merge branch 'master' of ssh://brick.kernel.dk/data/git/fio
[fio.git] / init.c
1 /*
2  * This file contains job initialization and setup functions.
3  */
4 #include <stdio.h>
5 #include <stdlib.h>
6 #include <unistd.h>
7 #include <fcntl.h>
8 #include <ctype.h>
9 #include <string.h>
10 #include <errno.h>
11 #include <getopt.h>
12 #include <assert.h>
13 #include <sys/ipc.h>
14 #include <sys/shm.h>
15 #include <sys/types.h>
16 #include <sys/stat.h>
17
18 #include "fio.h"
19 #include "parse.h"
20
21 #define FIO_RANDSEED            (0xb1899bedUL)
22
23 #define td_var_offset(var)      ((size_t) &((struct thread_data *)0)->var)
24
25 static int str_mem_cb(void *, const char *);
26 static int str_lockmem_cb(void *, unsigned long *);
27 #ifdef FIO_HAVE_IOPRIO
28 static int str_prio_cb(void *, unsigned int *);
29 static int str_prioclass_cb(void *, unsigned int *);
30 #endif
31 static int str_exitall_cb(void);
32 static int str_cpumask_cb(void *, unsigned int *);
33 static int str_fst_cb(void *, const char *);
34 static int str_filename_cb(void *, const char *);
35 static int str_directory_cb(void *, const char *);
36 static int str_opendir_cb(void *, const char *);
37
38 #define __stringify_1(x)        #x
39 #define __stringify(x)          __stringify_1(x)
40
41 /*
42  * Map of job/command line options
43  */
44 static struct fio_option options[] = {
45         {
46                 .name   = "description",
47                 .type   = FIO_OPT_STR_STORE,
48                 .off1   = td_var_offset(description),
49                 .help   = "Text job description",
50         },
51         {
52                 .name   = "name",
53                 .type   = FIO_OPT_STR_STORE,
54                 .off1   = td_var_offset(name),
55                 .help   = "Name of this job",
56         },
57         {
58                 .name   = "directory",
59                 .type   = FIO_OPT_STR_STORE,
60                 .off1   = td_var_offset(directory),
61                 .cb     = str_directory_cb,
62                 .help   = "Directory to store files in",
63         },
64         {
65                 .name   = "filename",
66                 .type   = FIO_OPT_STR_STORE,
67                 .off1   = td_var_offset(filename),
68                 .cb     = str_filename_cb,
69                 .help   = "File(s) to use for the workload",
70         },
71         {
72                 .name   = "opendir",
73                 .type   = FIO_OPT_STR_STORE,
74                 .off1   = td_var_offset(opendir),
75                 .cb     = str_opendir_cb,
76                 .help   = "Recursively add files from this directory and down",
77         },
78         {
79                 .name   = "rw",
80                 .type   = FIO_OPT_STR,
81                 .off1   = td_var_offset(td_ddir),
82                 .help   = "IO direction",
83                 .def    = "read",
84                 .posval = {
85                           { .ival = "read",
86                             .oval = TD_DDIR_READ,
87                             .help = "Sequential read",
88                           },
89                           { .ival = "write",
90                             .oval = TD_DDIR_WRITE,
91                             .help = "Sequential write",
92                           },
93                           { .ival = "randread",
94                             .oval = TD_DDIR_RANDREAD,
95                             .help = "Random read",
96                           },
97                           { .ival = "randwrite",
98                             .oval = TD_DDIR_RANDWRITE,
99                             .help = "Random write",
100                           },
101                           { .ival = "rw",
102                             .oval = TD_DDIR_RW,
103                             .help = "Sequential read and write mix",
104                           },
105                           { .ival = "randrw",
106                             .oval = TD_DDIR_RANDRW,
107                             .help = "Random read and write mix"
108                           },
109                 },
110         },
111         {
112                 .name   = "ioengine",
113                 .type   = FIO_OPT_STR_STORE,
114                 .off1   = td_var_offset(ioengine),
115                 .help   = "IO engine to use",
116                 .def    = "sync",
117                 .posval = {
118                           { .ival = "sync",
119                             .help = "Use read/write",
120                           },
121 #ifdef FIO_HAVE_LIBAIO
122                           { .ival = "libaio",
123                             .help = "Linux native asynchronous IO",
124                           },
125 #endif
126 #ifdef FIO_HAVE_POSIXAIO
127                           { .ival = "posixaio",
128                             .help = "POSIX asynchronous IO",
129                           },
130 #endif
131                           { .ival = "mmap",
132                             .help = "Memory mapped IO",
133                           },
134 #ifdef FIO_HAVE_SPLICE
135                           { .ival = "splice",
136                             .help = "splice/vmsplice based IO",
137                           },
138 #endif
139 #ifdef FIO_HAVE_SGIO
140                           { .ival = "sg",
141                             .help = "SCSI generic v3 IO",
142                           },
143 #endif
144                           { .ival = "null",
145                             .help = "Testing engine (no data transfer)",
146                           },
147                           { .ival = "net",
148                             .help = "Network IO",
149                           },
150 #ifdef FIO_HAVE_SYSLET
151                           { .ival = "syslet-rw",
152                             .help = "syslet enabled async pread/pwrite IO",
153                           },
154 #endif
155                           { .ival = "cpuio",
156                             .help = "CPU cycler burner engine",
157                           },
158                           { .ival = "external",
159                             .help = "Load external engine (append name)",
160                           },
161                 },
162         },
163         {
164                 .name   = "iodepth",
165                 .type   = FIO_OPT_INT,
166                 .off1   = td_var_offset(iodepth),
167                 .help   = "Amount of IO buffers to keep in flight",
168                 .def    = "1",
169         },
170         {
171                 .name   = "iodepth_batch",
172                 .type   = FIO_OPT_INT,
173                 .off1   = td_var_offset(iodepth_batch),
174                 .help   = "Number of IO to submit in one go",
175         },
176         {
177                 .name   = "iodepth_low",
178                 .type   = FIO_OPT_INT,
179                 .off1   = td_var_offset(iodepth_low),
180                 .help   = "Low water mark for queuing depth",
181         },
182         {
183                 .name   = "size",
184                 .type   = FIO_OPT_STR_VAL,
185                 .off1   = td_var_offset(total_file_size),
186                 .help   = "Size of device or file",
187         },
188         {
189                 .name   = "bs",
190                 .type   = FIO_OPT_STR_VAL_INT,
191                 .off1   = td_var_offset(bs[DDIR_READ]),
192                 .off2   = td_var_offset(bs[DDIR_WRITE]),
193                 .help   = "Block size unit",
194                 .def    = "4k",
195         },
196         {
197                 .name   = "bsrange",
198                 .type   = FIO_OPT_RANGE,
199                 .off1   = td_var_offset(min_bs[DDIR_READ]),
200                 .off2   = td_var_offset(max_bs[DDIR_READ]),
201                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
202                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
203                 .help   = "Set block size range (in more detail than bs)",
204         },
205         {
206                 .name   = "bs_unaligned",
207                 .type   = FIO_OPT_STR_SET,
208                 .off1   = td_var_offset(bs_unaligned),
209                 .help   = "Don't sector align IO buffer sizes",
210         },
211         {
212                 .name   = "offset",
213                 .type   = FIO_OPT_STR_VAL,
214                 .off1   = td_var_offset(start_offset),
215                 .help   = "Start IO from this offset",
216                 .def    = "0",
217         },
218         {
219                 .name   = "randrepeat",
220                 .type   = FIO_OPT_BOOL,
221                 .off1   = td_var_offset(rand_repeatable),
222                 .help   = "Use repeatable random IO pattern",
223                 .def    = "1",
224         },
225         {
226                 .name   = "norandommap",
227                 .type   = FIO_OPT_STR_SET,
228                 .off1   = td_var_offset(norandommap),
229                 .help   = "Accept potential duplicate random blocks",
230         },
231         {
232                 .name   = "nrfiles",
233                 .type   = FIO_OPT_INT,
234                 .off1   = td_var_offset(nr_files),
235                 .help   = "Split job workload between this number of files",
236                 .def    = "1",
237         },
238         {
239                 .name   = "openfiles",
240                 .type   = FIO_OPT_INT,
241                 .off1   = td_var_offset(open_files),
242                 .help   = "Number of files to keep open at the same time",
243         },
244         {
245                 .name   = "file_service_type",
246                 .type   = FIO_OPT_STR,
247                 .cb     = str_fst_cb,
248                 .off1   = td_var_offset(file_service_type),
249                 .help   = "How to select which file to service next",
250                 .def    = "roundrobin",
251                 .posval = {
252                           { .ival = "random",
253                             .oval = FIO_FSERVICE_RANDOM,
254                             .help = "Choose a file at random",
255                           },
256                           { .ival = "roundrobin",
257                             .oval = FIO_FSERVICE_RR,
258                             .help = "Round robin select files",
259                           },
260                 },
261         },
262         {
263                 .name   = "fsync",
264                 .type   = FIO_OPT_INT,
265                 .off1   = td_var_offset(fsync_blocks),
266                 .help   = "Issue fsync for writes every given number of blocks",
267                 .def    = "0",
268         },
269         {
270                 .name   = "direct",
271                 .type   = FIO_OPT_BOOL,
272                 .off1   = td_var_offset(odirect),
273                 .help   = "Use O_DIRECT IO (negates buffered)",
274                 .def    = "0",
275         },
276         {
277                 .name   = "buffered",
278                 .type   = FIO_OPT_BOOL,
279                 .off1   = td_var_offset(odirect),
280                 .neg    = 1,
281                 .help   = "Use buffered IO (negates direct)",
282                 .def    = "1",
283         },
284         {
285                 .name   = "overwrite",
286                 .type   = FIO_OPT_BOOL,
287                 .off1   = td_var_offset(overwrite),
288                 .help   = "When writing, set whether to overwrite current data",
289                 .def    = "0",
290         },
291         {
292                 .name   = "loops",
293                 .type   = FIO_OPT_INT,
294                 .off1   = td_var_offset(loops),
295                 .help   = "Number of times to run the job",
296                 .def    = "1",
297         },
298         {
299                 .name   = "numjobs",
300                 .type   = FIO_OPT_INT,
301                 .off1   = td_var_offset(numjobs),
302                 .help   = "Duplicate this job this many times",
303                 .def    = "1",
304         },
305         {
306                 .name   = "startdelay",
307                 .type   = FIO_OPT_INT,
308                 .off1   = td_var_offset(start_delay),
309                 .help   = "Only start job when this period has passed",
310                 .def    = "0",
311         },
312         {
313                 .name   = "runtime",
314                 .alias  = "timeout",
315                 .type   = FIO_OPT_STR_VAL_TIME,
316                 .off1   = td_var_offset(timeout),
317                 .help   = "Stop workload when this amount of time has passed",
318                 .def    = "0",
319         },
320         {
321                 .name   = "mem",
322                 .type   = FIO_OPT_STR,
323                 .cb     = str_mem_cb,
324                 .off1   = td_var_offset(mem_type),
325                 .help   = "Backing type for IO buffers",
326                 .def    = "malloc",
327                 .posval = {
328                           { .ival = "malloc",
329                             .oval = MEM_MALLOC,
330                             .help = "Use malloc(3) for IO buffers",
331                           },
332                           { .ival = "shm",
333                             .oval = MEM_SHM,
334                             .help = "Use shared memory segments for IO buffers",
335                           },
336 #ifdef FIO_HAVE_HUGETLB
337                           { .ival = "shmhuge",
338                             .oval = MEM_SHMHUGE,
339                             .help = "Like shm, but use huge pages",
340                           },
341 #endif
342                           { .ival = "mmap",
343                             .oval = MEM_MMAP,
344                             .help = "Use mmap(2) (file or anon) for IO buffers",
345                           },
346 #ifdef FIO_HAVE_HUGETLB
347                           { .ival = "mmaphuge",
348                             .oval = MEM_MMAPHUGE,
349                             .help = "Like mmap, but use huge pages",
350                           },
351 #endif
352                   },
353         },
354         {
355                 .name   = "verify",
356                 .type   = FIO_OPT_STR,
357                 .off1   = td_var_offset(verify),
358                 .help   = "Verify data written",
359                 .def    = "0",
360                 .posval = {
361                           { .ival = "0",
362                             .oval = VERIFY_NONE,
363                             .help = "Don't do IO verification",
364                           },
365                           { .ival = "crc32",
366                             .oval = VERIFY_CRC32,
367                             .help = "Use crc32 checksums for verification",
368                           },
369                           { .ival = "md5",
370                             .oval = VERIFY_MD5,
371                             .help = "Use md5 checksums for verification",
372                           },
373                 },
374         },
375         {
376                 .name   = "write_iolog",
377                 .type   = FIO_OPT_STR_STORE,
378                 .off1   = td_var_offset(write_iolog_file),
379                 .help   = "Store IO pattern to file",
380         },
381         {
382                 .name   = "read_iolog",
383                 .type   = FIO_OPT_STR_STORE,
384                 .off1   = td_var_offset(read_iolog_file),
385                 .help   = "Playback IO pattern from file",
386         },
387         {
388                 .name   = "exec_prerun",
389                 .type   = FIO_OPT_STR_STORE,
390                 .off1   = td_var_offset(exec_prerun),
391                 .help   = "Execute this file prior to running job",
392         },
393         {
394                 .name   = "exec_postrun",
395                 .type   = FIO_OPT_STR_STORE,
396                 .off1   = td_var_offset(exec_postrun),
397                 .help   = "Execute this file after running job",
398         },
399 #ifdef FIO_HAVE_IOSCHED_SWITCH
400         {
401                 .name   = "ioscheduler",
402                 .type   = FIO_OPT_STR_STORE,
403                 .off1   = td_var_offset(ioscheduler),
404                 .help   = "Use this IO scheduler on the backing device",
405         },
406 #endif
407         {
408                 .name   = "zonesize",
409                 .type   = FIO_OPT_STR_VAL,
410                 .off1   = td_var_offset(zone_size),
411                 .help   = "Give size of an IO zone",
412                 .def    = "0",
413         },
414         {
415                 .name   = "zoneskip",
416                 .type   = FIO_OPT_STR_VAL,
417                 .off1   = td_var_offset(zone_skip),
418                 .help   = "Space between IO zones",
419                 .def    = "0",
420         },
421         {
422                 .name   = "lockmem",
423                 .type   = FIO_OPT_STR_VAL,
424                 .cb     = str_lockmem_cb,
425                 .help   = "Lock down this amount of memory",
426                 .def    = "0",
427         },
428         {
429                 .name   = "rwmixcycle",
430                 .type   = FIO_OPT_INT,
431                 .off1   = td_var_offset(rwmixcycle),
432                 .help   = "Cycle period for mixed read/write workloads (msec)",
433                 .def    = "500",
434         },
435         {
436                 .name   = "rwmixread",
437                 .type   = FIO_OPT_INT,
438                 .off1   = td_var_offset(rwmixread),
439                 .maxval = 100,
440                 .help   = "Percentage of mixed workload that is reads",
441                 .def    = "50",
442         },
443         {
444                 .name   = "rwmixwrite",
445                 .type   = FIO_OPT_INT,
446                 .off1   = td_var_offset(rwmixwrite),
447                 .maxval = 100,
448                 .help   = "Percentage of mixed workload that is writes",
449                 .def    = "50",
450         },
451         {
452                 .name   = "nice",
453                 .type   = FIO_OPT_INT,
454                 .off1   = td_var_offset(nice),
455                 .help   = "Set job CPU nice value",
456                 .minval = -19,
457                 .maxval = 20,
458                 .def    = "0",
459         },
460 #ifdef FIO_HAVE_IOPRIO
461         {
462                 .name   = "prio",
463                 .type   = FIO_OPT_INT,
464                 .cb     = str_prio_cb,
465                 .help   = "Set job IO priority value",
466                 .minval = 0,
467                 .maxval = 7,
468         },
469         {
470                 .name   = "prioclass",
471                 .type   = FIO_OPT_INT,
472                 .cb     = str_prioclass_cb,
473                 .help   = "Set job IO priority class",
474                 .minval = 0,
475                 .maxval = 3,
476         },
477 #endif
478         {
479                 .name   = "thinktime",
480                 .type   = FIO_OPT_INT,
481                 .off1   = td_var_offset(thinktime),
482                 .help   = "Idle time between IO buffers (usec)",
483                 .def    = "0",
484         },
485         {
486                 .name   = "thinktime_spin",
487                 .type   = FIO_OPT_INT,
488                 .off1   = td_var_offset(thinktime_spin),
489                 .help   = "Start think time by spinning this amount (usec)",
490                 .def    = "0",
491         },
492         {
493                 .name   = "thinktime_blocks",
494                 .type   = FIO_OPT_INT,
495                 .off1   = td_var_offset(thinktime_blocks),
496                 .help   = "IO buffer period between 'thinktime'",
497                 .def    = "1",
498         },
499         {
500                 .name   = "rate",
501                 .type   = FIO_OPT_INT,
502                 .off1   = td_var_offset(rate),
503                 .help   = "Set bandwidth rate",
504         },
505         {
506                 .name   = "ratemin",
507                 .type   = FIO_OPT_INT,
508                 .off1   = td_var_offset(ratemin),
509                 .help   = "The bottom limit accepted",
510         },
511         {
512                 .name   = "ratecycle",
513                 .type   = FIO_OPT_INT,
514                 .off1   = td_var_offset(ratecycle),
515                 .help   = "Window average for rate limits (msec)",
516                 .def    = "1000",
517         },
518         {
519                 .name   = "invalidate",
520                 .type   = FIO_OPT_BOOL,
521                 .off1   = td_var_offset(invalidate_cache),
522                 .help   = "Invalidate buffer/page cache prior to running job",
523                 .def    = "1",
524         },
525         {
526                 .name   = "sync",
527                 .type   = FIO_OPT_BOOL,
528                 .off1   = td_var_offset(sync_io),
529                 .help   = "Use O_SYNC for buffered writes",
530                 .def    = "0",
531         },
532         {
533                 .name   = "bwavgtime",
534                 .type   = FIO_OPT_INT,
535                 .off1   = td_var_offset(bw_avg_time),
536                 .help   = "Time window over which to calculate bandwidth (msec)",
537                 .def    = "500",
538         },
539         {
540                 .name   = "create_serialize",
541                 .type   = FIO_OPT_BOOL,
542                 .off1   = td_var_offset(create_serialize),
543                 .help   = "Serialize creating of job files",
544                 .def    = "1",
545         },
546         {
547                 .name   = "create_fsync",
548                 .type   = FIO_OPT_BOOL,
549                 .off1   = td_var_offset(create_fsync),
550                 .help   = "Fsync file after creation",
551                 .def    = "1",
552         },
553         {
554                 .name   = "cpuload",
555                 .type   = FIO_OPT_INT,
556                 .off1   = td_var_offset(cpuload),
557                 .help   = "Use this percentage of CPU",
558         },
559         {
560                 .name   = "cpuchunks",
561                 .type   = FIO_OPT_INT,
562                 .off1   = td_var_offset(cpucycle),
563                 .help   = "Length of the CPU burn cycles (usecs)",
564                 .def    = "50000",
565         },
566 #ifdef FIO_HAVE_CPU_AFFINITY
567         {
568                 .name   = "cpumask",
569                 .type   = FIO_OPT_INT,
570                 .cb     = str_cpumask_cb,
571                 .help   = "CPU affinity mask",
572         },
573 #endif
574         {
575                 .name   = "end_fsync",
576                 .type   = FIO_OPT_BOOL,
577                 .off1   = td_var_offset(end_fsync),
578                 .help   = "Include fsync at the end of job",
579                 .def    = "0",
580         },
581         {
582                 .name   = "fsync_on_close",
583                 .type   = FIO_OPT_BOOL,
584                 .off1   = td_var_offset(fsync_on_close),
585                 .help   = "fsync files on close",
586                 .def    = "0",
587         },
588         {
589                 .name   = "unlink",
590                 .type   = FIO_OPT_BOOL,
591                 .off1   = td_var_offset(unlink),
592                 .help   = "Unlink created files after job has completed",
593                 .def    = "0",
594         },
595         {
596                 .name   = "exitall",
597                 .type   = FIO_OPT_STR_SET,
598                 .cb     = str_exitall_cb,
599                 .help   = "Terminate all jobs when one exits",
600         },
601         {
602                 .name   = "stonewall",
603                 .type   = FIO_OPT_STR_SET,
604                 .off1   = td_var_offset(stonewall),
605                 .help   = "Insert a hard barrier between this job and previous",
606         },
607         {
608                 .name   = "thread",
609                 .type   = FIO_OPT_STR_SET,
610                 .off1   = td_var_offset(use_thread),
611                 .help   = "Use threads instead of forks",
612         },
613         {
614                 .name   = "write_bw_log",
615                 .type   = FIO_OPT_STR_SET,
616                 .off1   = td_var_offset(write_bw_log),
617                 .help   = "Write log of bandwidth during run",
618         },
619         {
620                 .name   = "write_lat_log",
621                 .type   = FIO_OPT_STR_SET,
622                 .off1   = td_var_offset(write_lat_log),
623                 .help   = "Write log of latency during run",
624         },
625         {
626                 .name   = "hugepage-size",
627                 .type   = FIO_OPT_STR_VAL,
628                 .off1   = td_var_offset(hugepage_size),
629                 .help   = "When using hugepages, specify size of each page",
630                 .def    = __stringify(FIO_HUGE_PAGE),
631         },
632         {
633                 .name   = "group_reporting",
634                 .type   = FIO_OPT_STR_SET,
635                 .off1   = td_var_offset(group_reporting),
636                 .help   = "Do reporting on a per-group basis",
637         },
638         {
639                 .name = NULL,
640         },
641 };
642
643 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
644 #define FIO_CMD_OPTS    (16)
645 #define FIO_GETOPT_JOB  (0x89988998)
646
647 /*
648  * Command line options. These will contain the above, plus a few
649  * extra that only pertain to fio itself and not jobs.
650  */
651 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
652         {
653                 .name           = "output",
654                 .has_arg        = required_argument,
655                 .val            = 'o',
656         },
657         {
658                 .name           = "timeout",
659                 .has_arg        = required_argument,
660                 .val            = 't',
661         },
662         {
663                 .name           = "latency-log",
664                 .has_arg        = required_argument,
665                 .val            = 'l',
666         },
667         {
668                 .name           = "bandwidth-log",
669                 .has_arg        = required_argument,
670                 .val            = 'b',
671         },
672         {
673                 .name           = "minimal",
674                 .has_arg        = optional_argument,
675                 .val            = 'm',
676         },
677         {
678                 .name           = "version",
679                 .has_arg        = no_argument,
680                 .val            = 'v',
681         },
682         {
683                 .name           = "help",
684                 .has_arg        = no_argument,
685                 .val            = 'h',
686         },
687         {
688                 .name           = "cmdhelp",
689                 .has_arg        = optional_argument,
690                 .val            = 'c',
691         },
692         {
693                 .name           = NULL,
694         },
695 };
696
697 static int def_timeout = 0;
698
699 static char fio_version_string[] = "fio 1.14";
700
701 static char **ini_file;
702 static int max_jobs = MAX_JOBS;
703
704 struct thread_data def_thread;
705 struct thread_data *threads = NULL;
706
707 int exitall_on_terminate = 0;
708 int terse_output = 0;
709 unsigned long long mlock_size = 0;
710 FILE *f_out = NULL;
711 FILE *f_err = NULL;
712
713 static int write_lat_log = 0;
714 int write_bw_log = 0;
715
716 static int prev_group_jobs;
717
718 FILE *get_f_out()
719 {
720         return f_out;
721 }
722
723 FILE *get_f_err()
724 {
725         return f_err;
726 }
727
728 /*
729  * Return a free job structure.
730  */
731 static struct thread_data *get_new_job(int global, struct thread_data *parent)
732 {
733         struct thread_data *td;
734
735         if (global)
736                 return &def_thread;
737         if (thread_number >= max_jobs)
738                 return NULL;
739
740         td = &threads[thread_number++];
741         *td = *parent;
742
743         td->thread_number = thread_number;
744         return td;
745 }
746
747 static void put_job(struct thread_data *td)
748 {
749         if (td == &def_thread)
750                 return;
751
752         if (td->error)
753                 fprintf(f_out, "fio: %s\n", td->verror);
754
755         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
756         thread_number--;
757 }
758
759 /*
760  * Lazy way of fixing up options that depend on each other. We could also
761  * define option callback handlers, but this is easier.
762  */
763 static void fixup_options(struct thread_data *td)
764 {
765         if (!td->rwmixread && td->rwmixwrite)
766                 td->rwmixread = 100 - td->rwmixwrite;
767
768         if (td->write_iolog_file && td->read_iolog_file) {
769                 log_err("fio: read iolog overrides write_iolog\n");
770                 free(td->write_iolog_file);
771                 td->write_iolog_file = NULL;
772         }
773
774         if (td->io_ops->flags & FIO_SYNCIO)
775                 td->iodepth = 1;
776         else {
777                 if (!td->iodepth)
778                         td->iodepth = td->open_files;
779         }
780
781         /*
782          * only really works for sequential io for now, and with 1 file
783          */
784         if (td->zone_size && td_random(td) && td->open_files == 1)
785                 td->zone_size = 0;
786
787         /*
788          * Reads can do overwrites, we always need to pre-create the file
789          */
790         if (td_read(td) || td_rw(td))
791                 td->overwrite = 1;
792
793         if (!td->min_bs[DDIR_READ])
794                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
795         if (!td->max_bs[DDIR_READ])
796                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
797         if (!td->min_bs[DDIR_WRITE])
798                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
799         if (!td->max_bs[DDIR_WRITE])
800                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
801
802         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
803
804         if (td_read(td) && !td_rw(td))
805                 td->verify = 0;
806
807         if (td->norandommap && td->verify != VERIFY_NONE) {
808                 log_err("fio: norandommap given, verify disabled\n");
809                 td->verify = VERIFY_NONE;
810         }
811         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
812                 log_err("fio: bs_unaligned may not work with raw io\n");
813
814         /*
815          * thinktime_spin must be less than thinktime
816          */
817         if (td->thinktime_spin > td->thinktime)
818                 td->thinktime_spin = td->thinktime;
819
820         /*
821          * The low water mark cannot be bigger than the iodepth
822          */
823         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
824                 /*
825                  * syslet work around - if the workload is sequential,
826                  * we want to let the queue drain all the way down to
827                  * avoid seeking between async threads
828                  */
829                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
830                         td->iodepth_low = 1;
831                 else
832                         td->iodepth_low = td->iodepth;
833         }
834
835         /*
836          * If batch number isn't set, default to the same as iodepth
837          */
838         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
839                 td->iodepth_batch = td->iodepth;
840
841         if (td->nr_files > td->files_index)
842                 td->nr_files = td->files_index;
843
844         if (td->open_files > td->nr_files || !td->open_files)
845                 td->open_files = td->nr_files;
846 }
847
848 /*
849  * This function leaks the buffer
850  */
851 static char *to_kmg(unsigned int val)
852 {
853         char *buf = malloc(32);
854         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
855         char *p = post;
856
857         do {
858                 if (val & 1023)
859                         break;
860
861                 val >>= 10;
862                 p++;
863         } while (*p);
864
865         snprintf(buf, 31, "%u%c", val, *p);
866         return buf;
867 }
868
869 /* External engines are specified by "external:name.o") */
870 static const char *get_engine_name(const char *str)
871 {
872         char *p = strstr(str, ":");
873
874         if (!p)
875                 return str;
876
877         p++;
878         strip_blank_front(&p);
879         strip_blank_end(p);
880         return p;
881 }
882
883 /*
884  * Adds a job to the list of things todo. Sanitizes the various options
885  * to make sure we don't have conflicts, and initializes various
886  * members of td.
887  */
888 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
889 {
890         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
891                                    "randread", "randwrite", "randrw" };
892         unsigned int i;
893         struct fio_file *f;
894         const char *engine;
895         char fname[PATH_MAX];
896         int numjobs;
897
898         /*
899          * the def_thread is just for options, it's not a real job
900          */
901         if (td == &def_thread)
902                 return 0;
903
904         engine = get_engine_name(td->ioengine);
905         td->io_ops = load_ioengine(td, engine);
906         if (!td->io_ops) {
907                 log_err("fio: failed to load engine %s\n", engine);
908                 return 1;
909         }
910
911         if (td->use_thread)
912                 nr_thread++;
913         else
914                 nr_process++;
915
916         if (td->odirect)
917                 td->io_ops->flags |= FIO_RAWIO;
918
919         if (!td->filename && !td->files_index) {
920                 td->filename = strdup(jobname);
921
922                 if (td->nr_files == 1)
923                         add_file(td, td->filename);
924                 else {
925                         for (i = 0; i < td->nr_files; i++) {
926                                 sprintf(fname, "%s.%d.%d", td->filename, td->thread_number, i);
927                                 add_file(td, fname);
928                         }
929                 }
930         }
931
932         fixup_options(td);
933
934         for_each_file(td, f, i) {
935                 if (td->directory && f->filetype == FIO_TYPE_FILE) {
936                         sprintf(fname, "%s/%s", td->directory, f->file_name);
937                         f->file_name = strdup(fname);
938                 }
939         }
940                 
941         td->mutex = fio_sem_init(0);
942
943         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
944         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
945         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
946
947         if ((td->stonewall || td->numjobs > 1) && prev_group_jobs) {
948                 prev_group_jobs = 0;
949                 groupid++;
950         }
951
952         td->groupid = groupid;
953         prev_group_jobs++;
954
955         if (setup_rate(td))
956                 goto err;
957
958         if (td->write_lat_log) {
959                 setup_log(&td->ts.slat_log);
960                 setup_log(&td->ts.clat_log);
961         }
962         if (td->write_bw_log)
963                 setup_log(&td->ts.bw_log);
964
965         if (!td->name)
966                 td->name = strdup(jobname);
967
968         if (!terse_output) {
969                 if (!job_add_num) {
970                         if (!strcmp(td->io_ops->name, "cpuio"))
971                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
972                         else {
973                                 char *c1, *c2, *c3, *c4;
974
975                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
976                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
977                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
978                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
979
980                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
981
982                                 free(c1);
983                                 free(c2);
984                                 free(c3);
985                                 free(c4);
986                         }
987                 } else if (job_add_num == 1)
988                         fprintf(f_out, "...\n");
989         }
990
991         /*
992          * recurse add identical jobs, clear numjobs and stonewall options
993          * as they don't apply to sub-jobs
994          */
995         numjobs = td->numjobs;
996         while (--numjobs) {
997                 struct thread_data *td_new = get_new_job(0, td);
998
999                 if (!td_new)
1000                         goto err;
1001
1002                 td_new->numjobs = 1;
1003                 td_new->stonewall = 0;
1004                 job_add_num = numjobs - 1;
1005
1006                 if (add_job(td_new, jobname, job_add_num))
1007                         goto err;
1008         }
1009
1010         if (td->numjobs > 1) {
1011                 groupid++;
1012                 prev_group_jobs = 0;
1013         }
1014
1015         return 0;
1016 err:
1017         put_job(td);
1018         return -1;
1019 }
1020
1021 /*
1022  * Initialize the various random states we need (random io, block size ranges,
1023  * read/write mix, etc).
1024  */
1025 int init_random_state(struct thread_data *td)
1026 {
1027         unsigned long seeds[5];
1028         int fd, num_maps, blocks;
1029         struct fio_file *f;
1030         unsigned int i;
1031
1032         if (td->io_ops->flags & FIO_DISKLESSIO)
1033                 return 0;
1034
1035         fd = open("/dev/urandom", O_RDONLY);
1036         if (fd == -1) {
1037                 td_verror(td, errno, "open");
1038                 return 1;
1039         }
1040
1041         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
1042                 td_verror(td, EIO, "read");
1043                 close(fd);
1044                 return 1;
1045         }
1046
1047         close(fd);
1048
1049         os_random_seed(seeds[0], &td->bsrange_state);
1050         os_random_seed(seeds[1], &td->verify_state);
1051         os_random_seed(seeds[2], &td->rwmix_state);
1052
1053         if (td->file_service_type == FIO_FSERVICE_RANDOM)
1054                 os_random_seed(seeds[3], &td->next_file_state);
1055
1056         if (!td_random(td))
1057                 return 0;
1058
1059         if (td->rand_repeatable)
1060                 seeds[4] = FIO_RANDSEED * td->thread_number;
1061
1062         if (!td->norandommap) {
1063                 for_each_file(td, f, i) {
1064                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
1065                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
1066                         f->file_map = malloc(num_maps * sizeof(long));
1067                         if (!f->file_map) {
1068                                 log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
1069                                 return 1;
1070                         }
1071                         f->num_maps = num_maps;
1072                         memset(f->file_map, 0, num_maps * sizeof(long));
1073                 }
1074         }
1075
1076         os_random_seed(seeds[4], &td->random_state);
1077         return 0;
1078 }
1079
1080 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
1081 {
1082 #ifdef FIO_HAVE_CPU_AFFINITY
1083         unsigned int i;
1084
1085         CPU_ZERO(&cpumask);
1086
1087         for (i = 0; i < sizeof(int) * 8; i++) {
1088                 if ((1 << i) & cpu)
1089                         CPU_SET(i, &cpumask);
1090         }
1091 #endif
1092 }
1093
1094 static int is_empty_or_comment(char *line)
1095 {
1096         unsigned int i;
1097
1098         for (i = 0; i < strlen(line); i++) {
1099                 if (line[i] == ';')
1100                         return 1;
1101                 if (line[i] == '#')
1102                         return 1;
1103                 if (!isspace(line[i]) && !iscntrl(line[i]))
1104                         return 0;
1105         }
1106
1107         return 1;
1108 }
1109
1110 /*
1111  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1112  */
1113 static char *get_opt_postfix(const char *str)
1114 {
1115         char *p = strstr(str, ":");
1116
1117         if (!p)
1118                 return NULL;
1119
1120         p++;
1121         strip_blank_front(&p);
1122         strip_blank_end(p);
1123         return strdup(p);
1124 }
1125
1126 static int str_mem_cb(void *data, const char *mem)
1127 {
1128         struct thread_data *td = data;
1129
1130         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1131                 td->mmapfile = get_opt_postfix(mem);
1132                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1133                         log_err("fio: mmaphuge:/path/to/file\n");
1134                         return 1;
1135                 }
1136         }
1137
1138         return 0;
1139 }
1140
1141 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1142 {
1143         mlock_size = *val;
1144         return 0;
1145 }
1146
1147 #ifdef FIO_HAVE_IOPRIO
1148 static int str_prioclass_cb(void *data, unsigned int *val)
1149 {
1150         struct thread_data *td = data;
1151
1152         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1153         return 0;
1154 }
1155
1156 static int str_prio_cb(void *data, unsigned int *val)
1157 {
1158         struct thread_data *td = data;
1159
1160         td->ioprio |= *val;
1161         return 0;
1162 }
1163 #endif
1164
1165 static int str_exitall_cb(void)
1166 {
1167         exitall_on_terminate = 1;
1168         return 0;
1169 }
1170
1171 static int str_cpumask_cb(void *data, unsigned int *val)
1172 {
1173         struct thread_data *td = data;
1174
1175         fill_cpu_mask(td->cpumask, *val);
1176         return 0;
1177 }
1178
1179 static int str_fst_cb(void *data, const char *str)
1180 {
1181         struct thread_data *td = data;
1182         char *nr = get_opt_postfix(str);
1183
1184         td->file_service_nr = 1;
1185         if (nr)
1186                 td->file_service_nr = atoi(nr);
1187
1188         return 0;
1189 }
1190
1191 static int str_filename_cb(void *data, const char *input)
1192 {
1193         struct thread_data *td = data;
1194         char *fname, *str, *p;
1195
1196         p = str = strdup(input);
1197
1198         strip_blank_front(&str);
1199         strip_blank_end(str);
1200
1201         if (!td->files_index)
1202                 td->nr_files = 0;
1203
1204         while ((fname = strsep(&str, ":")) != NULL) {
1205                 if (!strlen(fname))
1206                         break;
1207                 add_file(td, fname);
1208                 td->nr_files++;
1209         }
1210
1211         free(p);
1212         return 0;
1213 }
1214
1215 static int str_directory_cb(void *data, const char fio_unused *str)
1216 {
1217         struct thread_data *td = data;
1218         struct stat sb;
1219
1220         if (lstat(td->directory, &sb) < 0) {
1221                 log_err("fio: %s is not a directory\n", td->directory);
1222                 td_verror(td, errno, "lstat");
1223                 return 1;
1224         }
1225         if (!S_ISDIR(sb.st_mode)) {
1226                 log_err("fio: %s is not a directory\n", td->directory);
1227                 return 1;
1228         }
1229
1230         return 0;
1231 }
1232
1233 static int str_opendir_cb(void *data, const char fio_unused *str)
1234 {
1235         struct thread_data *td = data;
1236
1237         if (!td->files_index)
1238                 td->nr_files = 0;
1239
1240         return add_dir_files(td, td->opendir);
1241 }
1242
1243 /*
1244  * This is our [ini] type file parser.
1245  */
1246 static int parse_jobs_ini(char *file, int stonewall_flag)
1247 {
1248         unsigned int global;
1249         struct thread_data *td;
1250         char *string, *name;
1251         fpos_t off;
1252         FILE *f;
1253         char *p;
1254         int ret = 0, stonewall;
1255
1256         f = fopen(file, "r");
1257         if (!f) {
1258                 perror("fopen job file");
1259                 return 1;
1260         }
1261
1262         string = malloc(4096);
1263         name = malloc(256);
1264         memset(name, 0, 256);
1265
1266         stonewall = stonewall_flag;
1267         do {
1268                 p = fgets(string, 4095, f);
1269                 if (!p)
1270                         break;
1271                 if (is_empty_or_comment(p))
1272                         continue;
1273                 if (sscanf(p, "[%255s]", name) != 1)
1274                         continue;
1275
1276                 global = !strncmp(name, "global", 6);
1277
1278                 name[strlen(name) - 1] = '\0';
1279
1280                 td = get_new_job(global, &def_thread);
1281                 if (!td) {
1282                         ret = 1;
1283                         break;
1284                 }
1285
1286                 /*
1287                  * Seperate multiple job files by a stonewall
1288                  */
1289                 if (!global && stonewall) {
1290                         td->stonewall = stonewall;
1291                         stonewall = 0;
1292                 }
1293
1294                 fgetpos(f, &off);
1295                 while ((p = fgets(string, 4096, f)) != NULL) {
1296                         if (is_empty_or_comment(p))
1297                                 continue;
1298
1299                         strip_blank_front(&p);
1300
1301                         if (p[0] == '[')
1302                                 break;
1303
1304                         strip_blank_end(p);
1305
1306                         fgetpos(f, &off);
1307
1308                         /*
1309                          * Don't break here, continue parsing options so we
1310                          * dump all the bad ones. Makes trial/error fixups
1311                          * easier on the user.
1312                          */
1313                         ret |= parse_option(p, options, td);
1314                 }
1315
1316                 if (!ret) {
1317                         fsetpos(f, &off);
1318                         ret = add_job(td, name, 0);
1319                 } else {
1320                         log_err("fio: job %s dropped\n", name);
1321                         put_job(td);
1322                 }
1323         } while (!ret);
1324
1325         free(string);
1326         free(name);
1327         fclose(f);
1328         return ret;
1329 }
1330
1331 static int fill_def_thread(void)
1332 {
1333         memset(&def_thread, 0, sizeof(def_thread));
1334
1335         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1336                 perror("sched_getaffinity");
1337                 return 1;
1338         }
1339
1340         /*
1341          * fill default options
1342          */
1343         fill_default_options(&def_thread, options);
1344
1345         def_thread.timeout = def_timeout;
1346         def_thread.write_bw_log = write_bw_log;
1347         def_thread.write_lat_log = write_lat_log;
1348
1349 #ifdef FIO_HAVE_DISK_UTIL
1350         def_thread.do_disk_util = 1;
1351 #endif
1352
1353         return 0;
1354 }
1355
1356 static void usage(void)
1357 {
1358         printf("%s\n", fio_version_string);
1359         printf("\t--output\tWrite output to file\n");
1360         printf("\t--timeout\tRuntime in seconds\n");
1361         printf("\t--latency-log\tGenerate per-job latency logs\n");
1362         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1363         printf("\t--minimal\tMinimal (terse) output\n");
1364         printf("\t--version\tPrint version info and exit\n");
1365         printf("\t--help\t\tPrint this page\n");
1366         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1367 }
1368
1369 static int parse_cmd_line(int argc, char *argv[])
1370 {
1371         struct thread_data *td = NULL;
1372         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1373
1374         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1375                 switch (c) {
1376                 case 't':
1377                         def_timeout = atoi(optarg);
1378                         break;
1379                 case 'l':
1380                         write_lat_log = 1;
1381                         break;
1382                 case 'w':
1383                         write_bw_log = 1;
1384                         break;
1385                 case 'o':
1386                         f_out = fopen(optarg, "w+");
1387                         if (!f_out) {
1388                                 perror("fopen output");
1389                                 exit(1);
1390                         }
1391                         f_err = f_out;
1392                         break;
1393                 case 'm':
1394                         terse_output = 1;
1395                         break;
1396                 case 'h':
1397                         usage();
1398                         exit(0);
1399                 case 'c':
1400                         ret = show_cmd_help(options, optarg);
1401                         exit(ret);
1402                 case 'v':
1403                         printf("%s\n", fio_version_string);
1404                         exit(0);
1405                 case FIO_GETOPT_JOB: {
1406                         const char *opt = long_options[lidx].name;
1407                         char *val = optarg;
1408
1409                         if (!strncmp(opt, "name", 4) && td) {
1410                                 ret = add_job(td, td->name ?: "fio", 0);
1411                                 if (ret) {
1412                                         put_job(td);
1413                                         return 0;
1414                                 }
1415                                 td = NULL;
1416                         }
1417                         if (!td) {
1418                                 int global = !strncmp(val, "global", 6);
1419
1420                                 td = get_new_job(global, &def_thread);
1421                                 if (!td)
1422                                         return 0;
1423                         }
1424
1425                         ret = parse_cmd_option(opt, val, options, td);
1426                         if (ret)
1427                                 dont_add_job = 1;
1428                         break;
1429                 }
1430                 default:
1431                         break;
1432                 }
1433         }
1434
1435         if (td) {
1436                 if (dont_add_job)
1437                         put_job(td);
1438                 else {
1439                         ret = add_job(td, td->name ?: "fio", 0);
1440                         if (ret)
1441                                 put_job(td);
1442                 }
1443         }
1444
1445         while (optind < argc) {
1446                 ini_idx++;
1447                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1448                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1449                 optind++;
1450         }
1451
1452         return ini_idx;
1453 }
1454
1455 static void free_shm(void)
1456 {
1457         struct shmid_ds sbuf;
1458
1459         if (threads) {
1460                 shmdt((void *) threads);
1461                 threads = NULL;
1462                 shmctl(shm_id, IPC_RMID, &sbuf);
1463         }
1464 }
1465
1466 /*
1467  * The thread area is shared between the main process and the job
1468  * threads/processes. So setup a shared memory segment that will hold
1469  * all the job info.
1470  */
1471 static int setup_thread_area(void)
1472 {
1473         /*
1474          * 1024 is too much on some machines, scale max_jobs if
1475          * we get a failure that looks like too large a shm segment
1476          */
1477         do {
1478                 size_t size = max_jobs * sizeof(struct thread_data);
1479
1480                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1481                 if (shm_id != -1)
1482                         break;
1483                 if (errno != EINVAL) {
1484                         perror("shmget");
1485                         break;
1486                 }
1487
1488                 max_jobs >>= 1;
1489         } while (max_jobs);
1490
1491         if (shm_id == -1)
1492                 return 1;
1493
1494         threads = shmat(shm_id, NULL, 0);
1495         if (threads == (void *) -1) {
1496                 perror("shmat");
1497                 return 1;
1498         }
1499
1500         atexit(free_shm);
1501         return 0;
1502 }
1503
1504 /*
1505  * Copy the fio options into the long options map, so we mirror
1506  * job and cmd line options.
1507  */
1508 static void dupe_job_options(void)
1509 {
1510         struct fio_option *o;
1511         unsigned int i;
1512
1513         i = 0;
1514         while (long_options[i].name)
1515                 i++;
1516
1517         o = &options[0];
1518         while (o->name) {
1519                 long_options[i].name = o->name;
1520                 long_options[i].val = FIO_GETOPT_JOB;
1521                 if (o->type == FIO_OPT_STR_SET)
1522                         long_options[i].has_arg = no_argument;
1523                 else
1524                         long_options[i].has_arg = required_argument;
1525
1526                 i++;
1527                 o++;
1528                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1529         }
1530 }
1531
1532 int parse_options(int argc, char *argv[])
1533 {
1534         int job_files, i;
1535
1536         f_out = stdout;
1537         f_err = stderr;
1538
1539         options_init(options);
1540
1541         dupe_job_options();
1542
1543         if (setup_thread_area())
1544                 return 1;
1545         if (fill_def_thread())
1546                 return 1;
1547
1548         job_files = parse_cmd_line(argc, argv);
1549
1550         for (i = 0; i < job_files; i++) {
1551                 if (fill_def_thread())
1552                         return 1;
1553                 if (parse_jobs_ini(ini_file[i], i))
1554                         return 1;
1555                 free(ini_file[i]);
1556         }
1557
1558         free(ini_file);
1559
1560         if (!thread_number) {
1561                 log_err("No jobs defined(s)\n");
1562                 return 1;
1563         }
1564
1565         return 0;
1566 }